This work studies the lot-streaming flexible job shop scheduling problem (LSFJSP). Conventional meta-heuristics can leverage domain knowledge to generate effective solutions, but often rely on manually designed rules or expensive per-instance search. Deep reinforcement learning (DRL) provides fast online scheduling. However, training a neural policy by RL is challenging, as early-stage policies often generate weak trajectories that provide limited learning signals, resulting in low sample efficiency. To address this issue, this paper proposes a search-to-learn memetic reinforcement learning (S2LMRL) framework for LSFJSP. In S2LMRL, memetic search refines RL-sampled trajectories and returns improved decisions as teacher signals for policy optimization. S2LMRL directly learns a neural scheduling policy as the main solution constructor, while memetic search acts as a training-time teacher. Moreover, the proposed S2LMRL framework is parallelizable, where GPU-based policy learning and CPU-based memetic refinement mutually enhance each other. Experimental results on small-scale training, mid-scale transfer, and cross-scale zero-shot testing show that S2LMRL consistently improves the learned policy over the PureRL baseline. The results also show that reinforcement feedback is necessary for effectively exploiting searched trajectories, and that S2LMRL provides a competitive quality–time tradeoff compared with strong dispatching rules, meta-heuristics, DRL baselines, and symbolic hyper-heuristics.
Ordinal classification problems are prevalent in real-world decision-making scenarios, where class labels are categorical yet inherently ordered. However, existing deep ordinal methods generally adopt a single ordinal modeling paradigm and treat all classes uniformly, making them insufficient for handling heterogeneous ordinal structures and class imbalance. To address these issues, this paper proposes a multi-head deep ordinal decision framework that integrates four representative ordinal paradigms as parallel output heads within a unified architecture, enabling the extraction of complementary ordinal characteristics. The Quadratic Weighted Kappa loss is adopted as a unified training objective, computed independently for each head and averaged to optimize the overall model. After training, the heterogeneous head outputs are combined through a class-aware fusion mechanism, formulated as a single-objective optimization problem that scalarizes the trade-off between prediction error and class-wise diversity. A quality- and similarity-driven differential evolution algorithm, equipped with quality-guided population initialization and similarity-aware mutation, is developed to search for optimal class-wise fusion weights. Experiments on ten public ordinal text classification datasets show that the proposed framework consistently outperforms state-of-the-art methods while maintaining robust predictive capability on minority classes.
Evolutionary algorithms (EAs) are effective for solving multimodal multiobjective optimization problem which optimal solutions subsets distributed regularly in the decision space. Existing EAs, however, are hard to achieve good diversity in the decision space due to their lack of thinking-in-space when solving deceptive multimodal multiobjective problems, and one manifestation of the deceptiveness is that global optima are composed of unconnected small optimal solution subsets that randomly distributed in the decision space with many local optima subsets mixed among. In this study, for enabling EAs to continuously search without trapping, a solving pathway is proposed that identifies information-rich historical data and constructs structured spatial representation based on high-quality evolutionary data. Furthermore, a moderate region-oriented solution set update methodology is proposed for the detection of ultrasmall optimal modalities, and two novel environmental selection and archive update mechanisms are designed. A Memory-aiding Search multimodal multiobjective Optimization EA, i.e., MS-multimodal multiobjective optimization EA (MMOEA), is instantiated. The experimental results on 27 deceptive instances and a real-world problem demonstrate MS-MMOEA's competitiveness over the state-of-the-art.
Surrogate-assisted evolutionary algorithms (SAEAs) have emerged as a promising way to reduce the computational cost of expensive constrained multiobjective optimisation problems (ECMOPs). However, under limited evaluation budgets, surrogate uncertainty near constraint boundaries may be amplified by inconsistent search decisions, thereby misleading the search and wasting expensive function evaluations. To address this issue, this paper proposes a probabilistic feasibility-guided (PF-Guided) framework that uses shared constraint-satisfaction probabilities to guide surrogate-population replacement and infill selection. Uncertainty-weighted global-local constraint modelling produces fused constraint-satisfaction probabilities for each candidate solution. The probability-vector constrained dominance principle (PV-CDP) uses the minimum and geometric mean of these probabilities across constraints for population replacement. Probability-feasible infill selection reuses the minimum constraint-satisfaction probability and feasibility threshold to select candidates for real evaluation. The PF-Guided framework is embedded in a two-stage SAEA, termed PF-SAEA, which combines unconstrained exploration with probability-guided constrained search. Experiments on 37 benchmark instances from the MW, DAS-CMOP, and LIRCMOP suites and six real-world engineering benchmark problems show that PF-SAEA achieves strong overall performance compared with seven representative peer algorithms and remains competitive on application-oriented ECMOPs.
Surrogate-assisted evolutionary algorithms (SAEAs) are widely used to tackle expensive optimization problems (EOPs), and their optimization performance largely depends on the effectiveness and robustness of the evolutionary search process. However, most existing SAEAs still rely on manually designed parameter and operator adaptation strategies, which are problem-dependent and require substantial expert knowledge. This paper proposes a large language model-assisted differential evolution optimizers automatic design (LLM-DOAD) framework to improve both the design efficiency of SAEAs and their performance in solving EOPs. Unlike redesigning surrogate model management, this paper focuses on automatically designing a differential evolution optimizer via LLMs. Specifically, problem-oriented prompts for EOPs guide LLMs to construct two essential modules, a parameter adaptation mechanism based on hybrid fitness improvement and dimension improvement and a multi-criteria and multi-stage operator adaptation mechanism. Moreover, a self-evolution learning approach is proposed to enhance the capability of LLMs in algorithm design and improve the quality of the generated algorithms. Finally, an LLM-based surrogate-assisted improved multi-operator differential evolution (LSIMD) algorithm is developed. Experiments on benchmark problems comprising 50, 100, and 200 dimensions show that LSIMD substantially outperforms several state-of-the-art manually designed algorithms, demonstrating the effectiveness and potential of LLM-DOAD for expensive optimization. LSIMD code is available at https://github.com/dabin64/LSIMD.git.
